$2.5B AI Chip Heist, Purpose-Built Robots, and the Future of American AI | This Week in AI Ep 6
TL;DR
Gecko’s bet is that robots are only valuable if they collect decisive data, not if they’re just cool hardware — Jake Loosararian said Gecko spent 13 years building purpose-built robots for refineries, shipyards, and power plants, feeding 500,000-600,000 critical assets into its Cantilever software to predict failures and eventually fix them.
Modular is attacking Nvidia lock-in by replacing the vendor software stack, not just wrapping it — Chris Lattner argued the real bottleneck isn’t a lack of chips but fragmented software, and said Modular now supports Nvidia, AMD, and Apple Silicon so developers can run across heterogeneous systems without rewriting everything.
The sleeper threat to Nvidia isn’t AMD — it’s Google’s TPU business if Google decides to really commercialize it — Lattner called Google the most underappreciated AI chip competitor because TPUs are already on their seventh generation and used at scale, but are still held back by poor ecosystem support and a lack of community adoption.
The $2.5B Nvidia chip smuggling story was framed as proof that AI compute is now a national-security race, not a normal tech market — reacting to reports of banned chips being routed into China with fake paperwork and altered markings, Loosararian said this is “an allout sprint” and a new kind of cold war over chips, energy, and industrial capacity.
Humanoid robots may get the headlines, but specialized robotics is where the near-term ROI is — after testing mobile platforms from the US, Europe, and China, Loosararian said general robots are still low-value in real industrial settings, while narrow systems for inspection, welding, and infrastructure maintenance can produce obvious, deterministic returns.
The hosts kept returning to a blunt labor thesis: AI will crush commodity knowledge work first, while trades and industrial jobs may become more attractive — they pointed to accounting rollups like Josh Kushner’s Thrive Holdings, argued the bottom 50% of CPA/lawyer work is becoming machine chores, and contrasted that with shortages in electricians, plumbers, radiologists, and robotics engineers.
The Breakdown
Gecko’s 13-year case for purpose-built robots
Jason opens with Jake Loosararian’s long arc: 13 years since building Gecko Robotics in a college dorm, and still very much not chasing the humanoid hype cycle. Jake’s core idea hasn’t changed — robots matter because they gather the right data about the “health of the built world,” basically “Minority Report for physical structures,” so you can predict a catastrophe before it happens.
Why Cantilever matters more than the robot body
The real product, Jake says, is Cantilever, the software layer that turns robot-collected inspection data into decisions: how to make a kilowatt cheaper, get more refinery output, or move a Navy ship out of dry dock faster. He draws a sharp line between robots-as-gadgets and robots tied to mission-critical outcomes, especially in settings where hallucinations can mean explosions, outages, or deaths.
Chris Lattner’s pitch: the AI world is stuck in chip-software duct tape
Chris Lattner explains Modular as the missing software layer between developers and a fragmented hardware world dominated by Nvidia, but increasingly including AMD and Apple Silicon. His diagnosis is blunt: CUDA, ROCm, and other vendor stacks are siloed because chip companies don’t want to play together, and the result is a mess of brittle compatibility issues built on “duct tape and bailing wire.”
The future is heterogeneous compute, and enterprises want choice
Lattner says Modular’s point is not just portability but letting different architectures actually work together — CPUs, GPUs, and custom AI accelerators in one system. He uses Nvidia’s upcoming Vera Rubin/Grock setup as the example: the world is moving toward mixed compute, and enterprises want the freedom to stay on Nvidia when it’s best but still shift workloads elsewhere without doubling their software headaches.
The chip shortage is real — and Google may be the most underrated threat to Nvidia
Asked whether the supply crunch is mostly hype, Chris says no: try buying 100 Blackwell nodes and you’ll run into long waits unless you’re one of the biggest players. His surprise ranking of Nvidia challengers starts with Google, whose TPUs are on generation seven and already powerful at scale, followed by AMD or possibly Amazon’s Tranium, but he says all of them are held back by one thing: terrible ecosystem accessibility compared with Nvidia’s decades of community-building around CUDA.
The $2.5B chip-heist story turns the conversation geopolitical fast
When Jason brings up the alleged smuggling of $2.5 billion in Nvidia chips into China, complete with fake paperwork and altered labels, Jake instantly zooms out to national security. For him, it’s evidence that AI infrastructure is an “allout sprint,” that chips and energy are now strategic assets, and that the US government should be spending much more aggressively and regulating with far less historical caution.
Self-driving and humanoids: cool demos, but where’s the actual ROI?
On autonomy, Chris says Waymo is the only serious scaled player right now and dismisses Tesla as not yet meaningfully autonomous until it files the right permits and expands beyond a tiny supervised Austin program. Jake is similarly skeptical of humanoid hype: he’s tested top walking robots from the US, Europe, and China and says they’re impressive on actuation and perception, but still weak on dexterity and, more importantly, not delivering enough economic value in industrial settings.
The final thesis: AI will commoditize white-collar chores, while industry gets rebuilt by tools that make hard jobs cool again
The back half gets surprisingly human. They talk about Josh Kushner’s Thrive rollup buying accounting firms, the idea that the bottom half of CPA and legal work is becoming machine labor, and the bigger opportunity in reviving manufacturing, mining, and energy jobs that have barely changed in 40 to 60 years. Jake’s line is the one that sticks: “Get robots everywhere. Get AI everywhere,” not just for productivity, but to make these sectors exciting enough that people actually want to build their lives there again.